View
1
Download
0
Category
Preview:
Citation preview
The ParLab Stack
Parallel Computing Laboratory
Sarah Bird
May 30, 2013
Bridging the Gap
Parallel Applications
Parallel Hardware
Parallel Software
IT industry Users
Easy to write correct programs that run efficiently on manycore
Integrated Software Stack Personal Health
Image Retrieval
Hearing, Music
Speech Parallel Browser
Motifs/Dwarfs
Sketching
Legacy Code
Schedulers Communication & Synch. Primitives
Efficiency Language Compilers
Legacy OS
Multicore/GPGPU
OS Libraries & Services
RAMP Manycore
Hypervisor
Co
rrec
tnes
s
Composition & Coordination Language (C&CL)
Parallel Libraries
Parallel Frameworks
Static Verification
Dynamic Checking
Debugging with Replay
Directed Testing
Autotuners
C&CL Compiler/Interpreter
Efficiency Languages
Type Systems
Hardware
• RAMP Gold
– Simulation
– 64 cores
– FPGA
RISC-V
Chip
FPGA
RAMP Gold
SPARC PTX
GPGPU
x86
Multicore
• RISC-V – Implementations
written in Chisel – Rocket
• 6 stage in-order
– Hwacha • 64 bit vector core
– FPGA • 2 Rocket
– 45 nm Chip • 1 Rocket, 1 Hwacha • 1 Ghz
• x86
• GPGPU
– Cuda
– OpenGL
Operating Systems
• Akaros
– Cloud OS
RISC-V
Chip
FPGA
RAMP Gold
SPARC PTX
GPGPU
x86
Multicore
• Tessellation
– Client OS
– Space-Time Partitioning
– Two-Level Scheduling
– QoS to Applications
– PACORA
• Linux
Akaros Tessellation Linux
Schedulers
• Lithe – Compose Parallel Runtimes
– Thread Building Blocks
– Open MP
RISC-V
Chip
FPGA
RAMP Gold
SPARC PTX
GPGPU
x86
Multicore
• PULSE – Framework to write schedulers
– Earliest Deadline-First
– Global Round Robin
Akaros Tessellation Linux
PULSE Lithe
OpenMP TBB EDF GRR
SEJITS
RISC-V
Chip
FPGA
RAMP Gold
SPARC PTX
GPGPU
x86
Multicore
Akaros Tessellation Linux
PULSE Lithe
OpenMP TBB EDF GRR
ASP Copperhead
PyCASP
TFJ
Applications
RISC-V
Chip
FPGA
RAMP Gold
SPARC PTX
GPGPU
x86
Multicore
Akaros Tessellation Linux
PULSE Lithe
OpenMP TBB EDF GRR
TFJ ASP Copperhead
Optical Flow Music
Synthesis
PyCASP
Parallel Browser
PARDORA
Why Create an Integrated Prototype?
• Encourages Collaboration • Prevents neglecting important
pieces of the problem • Uncover opportunities for
invention by seeing which side of an interface is the best place to satisfy a requirement
• Demonstrate the importance of design simplicity
• Enhance the education of the PhD students in areas beyond their own specialties
• Help with technology transfer by giving concrete examples of our ideas for our colleagues in industry
Forces for Integration • Design Compatibility
– Shared Space and Discussions – Symbiotic Designs – Example: Music and Tessellation
• Customized Support – In-house experts helping adapt
their design to your problem – Examples: Lithe and Tessellation CAA and Applications
• Motivating Applications – Exciting to show your research on
run a compelling application – Examples: Patterns and MRI BFS and RISC-V
Integrated Demos in ParLab History
Stack Redesign
• Productivity programs can create applications that require efficiency – Scale – Performance
• Easily target many platforms and features – Example: Vector units on RISC-V Chip
• Efficient performance predictibility – Interactivity and Responsiveness – Realtime Performance
What did rethinking the entire computing stack at once get us?
Integrated Demos
• Two Demos to show off the ParLab Stack
• Fun and compelling applications that require efficiency
• Easily target many platforms and features
• Interactivity and Realtime Performance
• Integration!
Music Exploration and Recommendation
• Better Pandora
• Audio Content Analysis Framework
• Parallel Browser Big Data Visualization
• Demonstrates Scale and Responsiveness
Music Recommendation and Exploration
RISC-V
RV FPGA
PTX
GPGPU
x86
Multicore
Tessellation
Lithe
TBB
ASP
PyCASP
Parallel Browser
PARDORA
RISC-V
RV FPGA
PTX
GPGPU
x86
Multicore
Tessellation
Lithe
TBB
ASP
PyCASP
Parallel Browser
PARDORA
Virtual Musical Instrument
• Musical Instrument using a camera
• Music Synthesis
• Vision Applications
• Demonstrate Realtime
Virtual Musical Instrument
PTX
GPGPU
x86
Multicore
Tess Linux
PULSE
GRR
TFJ
Optical Flow
Music Synthesis
RISC-V
FPGA
Chip
Virtual Musical Instrument
PTX
GPGPU
x86
Multicore
Tess Linux
PULSE
GRR
TFJ
Optical Flow
Music Synthesis
RISC-V
FPGA
Chip
Virtual Musical Instrument
PTX
GPGPU
x86
Multicore
Linux
PULSE
GRR
TFJ
Optical Flow
Music Synthesis
RISC-V
RV FPGA
RV Chip
Tess
Virtual Musical Instrument
PTX
GPGPU
x86
Multicore
Tess Linux
PULSE
GRR
TFJ
Optical Flow
Music Synthesis
RISC-V
RV Chip
RV Chip
Music Recommendation & Exploration Demo
Leo Meyerovich, Katya Gonina, Gage Eads, Eric Roman, Eric Battenberg, Henry Cook, Gerald Friedland
End of ParLab Celebration
May 30, 2013
21
Demo
22
System Overview
“Radiohead”
Client Server
Recommendation Engine 1K
clustered songs
23
1M songs
Layout Engine
sid_1, sid_2
“King Tubby”
SEJITS Parallel Browser
Tesselation
Client Architecture
24
Parallel JavaScript Libraries 1K
clustered songs
1M songs
Safari Web Browser
Parse (JSON)
Layout Render
viz.ftl Synthesizer
offline: compile-time
Music Recommendation Stack
ASP
PyCASP
OpenCL
GPGPU
Parallel Browser
PARDORA
Linux
CUDA / Cilk+
PTX
GPGPU
RISC-V
RV FPGA
x86
Multicore
Tessellation
Lithe
TBB
x86
Multicore
JavaScript
x86
Train a UBM*
Offline Phase
Online Phase
Find Closest Songs
Adapt UBM on query songs
Get potential neighbors using
Collaborative Filtering
* UBM = Universal Background Model 26
1M songs
1K clustered
songs
“Radiohead”
Server Architecture
SEJITS
GPU
CPU
FPGA
Tesselation
Music Recommendation Stack
ASP
PyCASP
OpenCL
GPGPU
Parallel Browser
PARDORA
Linux
CUDA / Cilk+
PTX
GPGPU
RISC-V
RV FPGA
x86
Multicore
Tessellation
Lithe
TBB
x86
Multicore
JavaScript
x86
Music Recommendation Stack
ASP
PyCASP
OpenCL
GPGPU
Parallel Browser
PARDORA
Tessellation
Lithe
TBB
x86
Multicore
RISC-V
RV FPGA
Linux
CUDA / Cilk+
PTX
GPGPU
x86
Multicore
Recommended